{
  "submission": "team5_HIPE-2026-v1.0-surprise-test-fr_run2.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team5",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          90,
          33,
          7
        ],
        [
          19,
          29,
          12
        ],
        [
          74,
          59,
          157
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 90,
          "pred_PROBABLE": 33,
          "pred_FALSE": 7
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 19,
          "pred_PROBABLE": 29,
          "pred_FALSE": 12
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 74,
          "pred_PROBABLE": 59,
          "pred_FALSE": 157
        }
      ],
      "accuracy": 0.575,
      "per_label": {
        "TRUE": {
          "precision": 0.4918,
          "recall": 0.6923,
          "f1": 0.5751,
          "support": 130,
          "true_positive": 90,
          "false_positive": 93,
          "false_negative": 40
        },
        "PROBABLE": {
          "precision": 0.2397,
          "recall": 0.4833,
          "f1": 0.3205,
          "support": 60,
          "true_positive": 29,
          "false_positive": 92,
          "false_negative": 31
        },
        "FALSE": {
          "precision": 0.892,
          "recall": 0.5414,
          "f1": 0.6738,
          "support": 290,
          "true_positive": 157,
          "false_positive": 19,
          "false_negative": 133
        }
      }
    }
  }
}